-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathloss.py
290 lines (210 loc) · 12.8 KB
/
loss.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
import torch
from utils import *
dtype = torch.cuda.FloatTensor
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
alpha, gamma0, gamma1 = 0.1, 0.8, 0.8
# batch_size while optimization is 1. changing this may increase efficiency.
def closure(cnn, preprocess, prior_nets, net_inputs, patch_num, imsize, map_idxs, counter, keys):
# global key_p0_t0, key_p0_t1, key_p1_t0, key_p1_t1
global alpha, gamma0, gamma1
key_p0_t0, key_p0_t1, key_p1_t0, key_p1_t1 = keys
patches = [prior_nets[i](net_inputs[i]) for i in range(patch_num)]
random_bg = torch.clamp(torch.zeros([1, 3, imsize, imsize]).normal_(mean=0.5, std=0.1).type(dtype), 0, 1).to(device)
random_bg_out = cnn(preprocess(random_bg)).to(device)
patch_positions = random_patch_placer(random_bg, patches)
single_patch_pics = []
single_patch_out = []
for i in range(patch_num):
pic = add_patches_img(random_bg, [patch_positions[i]])
single_patch_pics.append(pic)
single_patch_out.append(cnn(preprocess(pic)))
combined_patch_pic0 = add_patches_img(single_patch_pics[1].clone().detach(), [patch_positions[0]])
combined_patch_out0 = cnn(preprocess(combined_patch_pic0))
combined_patch_pic1 = add_patches_img(single_patch_pics[0].clone().detach(), [patch_positions[1]])
combined_patch_out1 = cnn(preprocess(combined_patch_pic1))
p0_t0 = single_patch_out[0][0, map_idxs[0]]
p0_t1 = (combined_patch_out0[0, map_idxs[1]] - single_patch_out[1][0, map_idxs[1]].clone().detach())**2
p1_t0 = single_patch_out[1][0, map_idxs[1]]
p1_t1 = (combined_patch_out1[0, map_idxs[0]] - single_patch_out[0][0, map_idxs[0]].clone().detach())**2
m = nn.Softmax(dim=1)
normalized_single_patch_out = [m(single_patch_out[i]).clone().detach() for i in range(patch_num)]
normalized_combined_patch_out0 = m(combined_patch_out0).clone().detach()
normalized_combined_patch_out1 = m(combined_patch_out1).clone().detach()
# TODO: We may want to adjust p{i}_t1 terms. Dropin 20 lower bound would make the explanation more persuasive and perhaps the optimization easier
with torch.no_grad():
normalized_p0_t0 = normalized_single_patch_out[0][0, map_idxs[0]]
normalized_p0_t1 = torch.tanh(abs(combined_patch_out0[0, map_idxs[1]] - single_patch_out[1][0, map_idxs[1]])
/min(abs(combined_patch_out0[0, map_idxs[1]].item()), abs(single_patch_out[1][0, map_idxs[1]]))).clone().detach()
normalized_p1_t0 = normalized_single_patch_out[1][0, map_idxs[1]]
normalized_p1_t1 = torch.tanh(abs(combined_patch_out1[0, map_idxs[0]] - single_patch_out[0][0, map_idxs[0]])
/min(abs(combined_patch_out1[0, map_idxs[0]].item()), abs(single_patch_out[0][0, map_idxs[0]]))).clone().detach()
if counter > 300:
key_p0_t0 = normalized_p0_t0 * alpha + (1 - alpha) * key_p0_t0
key_p1_t0 = normalized_p1_t0 * alpha + (1 - alpha) * key_p1_t0
key_p0_t1 = normalized_p0_t1 * alpha + (1 - alpha) * key_p0_t1
key_p1_t1 = normalized_p1_t1 * alpha + (1 - alpha) * key_p1_t1
if counter <= 300:
loss_patch0 = -5 * p0_t0
loss_patch1 = -5 * p1_t0
else:
gammas = [gamma0, gamma1]
loss_patch0 = -10 * FL(key_p0_t0, 1, gammas) * p0_t0 + 5 * FL(key_p0_t1, 0, gammas) * p0_t1
loss_patch1 = -10 * FL(key_p1_t0, 1, gammas) * p1_t0 + 5 * FL(key_p1_t1, 0, gammas) * p1_t1
# Back Prop
loss_patch0.backward()
loss_patch1.backward()
P0 = [normalized_p0_t0.item(), normalized_p1_t0.item()]
P1 = [normalized_p0_t1.item(), normalized_p1_t1.item()]
SCORES = [single_patch_out[i][0, map_idxs[i]].item() for i in range(patch_num)]
keys = key_p0_t0, key_p0_t1, key_p1_t0, key_p1_t1
return P0, P1, combined_patch_pic0, SCORES, patch_positions, random_bg, keys
def closure_null(cnn, preprocess, prior_nets, net_inputs, patch_num, imsize, map_idxs, counter, keys):
# global key_p0_t0, key_p0_t1, key_p1_t0, key_p1_t1
global alpha, gamma0, gamma1
key_p0_t0, key_p0_t1, key_p0_t2, key_p1_t0 = keys
patches = [prior_nets[i](net_inputs[i]) for i in range(patch_num)]
random_bg = torch.clamp(torch.zeros([1, 3, imsize, imsize]).normal_(mean=0.5, std=0.1).type(dtype), 0, 1).to(device)
random_bg_out = cnn(preprocess(random_bg)).to(device)
patch_positions = random_patch_placer(random_bg, patches)
single_patch_pics = []
single_patch_out = []
for i in range(patch_num):
pic = add_patches_img(random_bg, [patch_positions[i]])
single_patch_pics.append(pic)
single_patch_out.append(cnn(preprocess(pic)))
combined_patch_pic0 = add_patches_img(single_patch_pics[1].clone().detach(), [patch_positions[0]])
combined_patch_out0 = cnn(preprocess(combined_patch_pic0))
combined_patch_pic1 = add_patches_img(single_patch_pics[0].clone().detach(), [patch_positions[1]])
combined_patch_out1 = cnn(preprocess(combined_patch_pic1))
p0_t0 = single_patch_out[0][0, map_idxs[0]]
p0_t1 = (combined_patch_out0[0, map_idxs[1]] - single_patch_out[1][0, map_idxs[1]].clone().detach())**2
p0_t2 = (single_patch_out[0][0, map_idxs[1]] - random_bg_out[0, map_idxs[1]].clone().detach())**2
p1_t0 = single_patch_out[1][0, map_idxs[1]]
p1_t1 = (combined_patch_out1[0, map_idxs[0]] - single_patch_out[0][0, map_idxs[0]].clone().detach())**2
m = nn.Softmax(dim=1)
normalized_single_patch_out = [m(single_patch_out[i]) for i in range(patch_num)]
normalized_combined_patch_out0 = m(combined_patch_out0).clone().detach()
normalized_combined_patch_out1 = m(combined_patch_out1).clone().detach()
# TODO: We may want to adjust p{i}_t1 terms. Dropin 20 lower bound would make the explanation more persuasive and perhaps the optimization easier
normalized_p1_t0 = normalized_single_patch_out[1][0, map_idxs[1]]
with torch.no_grad():
normalized_p0_t0 = normalized_single_patch_out[0][0, map_idxs[0]]
normalized_p0_t1 = torch.tanh(abs(combined_patch_out0[0, map_idxs[1]] - single_patch_out[1][0, map_idxs[1]])
/min(abs(combined_patch_out0[0, map_idxs[1]].item()), abs(single_patch_out[1][0, map_idxs[1]]))).clone().detach()
normalized_p0_t2 = torch.tanh(abs(single_patch_out[0][0, map_idxs[1]] - random_bg_out[0, map_idxs[1]])
/min(abs(random_bg_out[0, map_idxs[1]].item()), abs(single_patch_out[0][0, map_idxs[1]]))).clone().detach()
if counter > 300:
key_p0_t0 = normalized_p0_t0 * alpha + (1 - alpha) * key_p0_t0
key_p0_t1 = normalized_p0_t1 * alpha + (1 - alpha) * key_p0_t1
key_p0_t2 = normalized_p0_t2 * alpha + (1 - alpha) * key_p0_t2
key_p1_t0 = normalized_p1_t0 * alpha + (1 - alpha) * key_p1_t0
if counter <= 300:
loss_patch0 = -5 * p0_t0
loss_patch1 = 5 * (normalized_p1_t0 - 0.95)**2
else:
gammas = [gamma0, gamma1]
loss_patch0 = -10 * FL(key_p0_t0, 1, gammas) * p0_t0 + 5 * FL(key_p0_t1, 0, gammas) * p0_t1 + 5 * FL(key_p0_t2, 0, gammas) * p0_t2
loss_patch1 = 5 * (normalized_p1_t0 - 0.95)**2
# Back Prop
loss_patch0.backward()
loss_patch1.backward()
P0 = [normalized_p0_t0.item(), normalized_p1_t0.item()]
P1 = [normalized_p0_t1.item(), normalized_p0_t2.item()]
SCORES = [single_patch_out[i][0, map_idxs[i]].item() for i in range(patch_num)]
keys = key_p0_t0, key_p0_t1, key_p0_t2, key_p1_t0
return P0, P1, combined_patch_pic0, SCORES, patch_positions, random_bg, keys
def closure_single_patch(cnn, preprocess, prior_nets, net_inputs, patch_num, imsize, map_idxs, counter, keys):
# global key_p0_t0, key_p0_t1
global alpha, gamma0, gamma1
key_p0_t0, key_p0_t1 = keys
patches = [prior_nets[i](net_inputs[i]) for i in range(patch_num)]
random_bg = torch.clamp(torch.zeros([1, 3, imsize, imsize]).normal_(mean=0.5, std=0.1).type(dtype), 0, 1).to(device)
random_bg_out = cnn(preprocess(random_bg)).to(device)
patch_positions = random_patch_placer(random_bg, patches)
single_patch_pics = []
single_patch_out = []
for i in range(patch_num):
pic = add_patches_img(random_bg, [patch_positions[i]])
single_patch_pics.append(pic)
single_patch_out.append(cnn(preprocess(pic)))
p0_t0 = single_patch_out[0][0, map_idxs[0]]
p0_t1 = (single_patch_out[0][0, map_idxs[1]] - random_bg_out[0, map_idxs[1]].clone().detach())**2
m = nn.Softmax(dim=1)
normalized_single_patch_out = [m(single_patch_out[i]).clone().detach() for i in range(patch_num)]
# TODO: We may want to adjust p{i}_t1 terms. Dropin 20 lower bound would make the explanation more persuasive and perhaps the optimization easier
with torch.no_grad():
normalized_p0_t0 = normalized_single_patch_out[0][0, map_idxs[0]]
normalized_p0_t1 = torch.tanh(abs(single_patch_out[0][0, map_idxs[1]] - random_bg_out[0, map_idxs[1]])
/min(abs(random_bg_out[0, map_idxs[1]].item()), abs(single_patch_out[0][0, map_idxs[1]]))).clone().detach()
if counter > 300:
key_p0_t0 = normalized_p0_t0 * alpha + (1 - alpha) * key_p0_t0
key_p0_t1 = normalized_p0_t1 * alpha + (1 - alpha) * key_p0_t1
if counter <= 300:
loss_patch0 = -5 * p0_t0
else:
gammas = [gamma0, gamma1]
loss_patch0 = -10 * FL(key_p0_t0, 1, gammas) * p0_t0 + 5 * FL(key_p0_t1, 0, gammas) * p0_t1
# Back Prop
loss_patch0.backward()
P0 = [normalized_p0_t0.item()]
P1 = [normalized_p0_t1.item()]
SCORES = [single_patch_out[i][0, map_idxs[i]].item() for i in range(patch_num)]
keys = key_p0_t0, key_p0_t1
return P0, P1, single_patch_pics[0], SCORES, patch_positions, random_bg, keys
def closure_repeated_patch(cnn, preprocess, prior_nets, net_inputs, patch_num, imsize, map_idxs, counter, keys):
# global key_p0_t0, key_p0_t1, key_p1_t0, key_p1_t1
global alpha, gamma0, gamma1
key_p0_t0, key_p0_t1, key_p1_t0, key_p1_t1 = keys
patches = [prior_nets[i](net_inputs[i]) for i in range(patch_num)]
map_idxs = [map_idxs[0], map_idxs[0]]
random_bg = torch.clamp(torch.zeros([1, 3, imsize, imsize]).normal_(mean=0.5, std=0.1).type(dtype), 0, 1).to(device)
random_bg_out = cnn(preprocess(random_bg)).to(device)
patch_positions = random_patch_placer(random_bg, patches)
patch_positions[0][:2] = [5, 5]
patch_positions[1][:2] = [133, 133]
single_patch_pics = []
single_patch_out = []
for i in range(patch_num):
pic = add_patches_img(random_bg, [patch_positions[i]])
single_patch_pics.append(pic)
single_patch_out.append(cnn(preprocess(pic)))
combined_patch_pic0 = add_patches_img(single_patch_pics[1].clone().detach(), [patch_positions[0]])
combined_patch_out0 = cnn(preprocess(combined_patch_pic0))
combined_patch_pic1 = add_patches_img(single_patch_pics[0].clone().detach(), [patch_positions[1]])
combined_patch_out1 = cnn(preprocess(combined_patch_pic1))
p0_t0 = single_patch_out[0][0, map_idxs[0]]
p0_t1 = (combined_patch_out0[0, map_idxs[1]] - single_patch_out[1][0, map_idxs[1]].clone().detach())**2
p1_t0 = single_patch_out[1][0, map_idxs[1]]
p1_t1 = (combined_patch_out1[0, map_idxs[0]] - single_patch_out[0][0, map_idxs[0]].clone().detach())**2
m = nn.Softmax(dim=1)
normalized_single_patch_out = [m(single_patch_out[i]) for i in range(patch_num)]
normalized_combined_patch_out0 = m(combined_patch_out0).clone().detach()
normalized_combined_patch_out1 = m(combined_patch_out1).clone().detach()
# TODO: We may want to adjust p{i}_t1 terms. Dropin 20 lower bound would make the explanation more persuasive and perhaps the optimization easier
with torch.no_grad():
normalized_p0_t0 = normalized_single_patch_out[0][0, map_idxs[0]]
normalized_p0_t1 = torch.tanh(abs(combined_patch_out0[0, map_idxs[1]] - single_patch_out[1][0, map_idxs[1]])
/min(abs(combined_patch_out0[0, map_idxs[1]].item()), abs(single_patch_out[1][0, map_idxs[1]]))).clone().detach()
normalized_p1_t0 = normalized_single_patch_out[1][0, map_idxs[1]]
normalized_p1_t1 = torch.tanh(abs(combined_patch_out1[0, map_idxs[0]] - single_patch_out[0][0, map_idxs[0]])
/min(abs(combined_patch_out1[0, map_idxs[0]].item()), abs(single_patch_out[0][0, map_idxs[0]]))).clone().detach()
if counter > 200:
key_p0_t0 = normalized_p0_t0 * alpha + (1 - alpha) * key_p0_t0
key_p1_t0 = normalized_p1_t0 * alpha + (1 - alpha) * key_p1_t0
key_p0_t1 = normalized_p0_t1 * alpha + (1 - alpha) * key_p0_t1
key_p1_t1 = normalized_p1_t1 * alpha + (1 - alpha) * key_p1_t1
if counter <= 200:
loss_patch0 = -5 * p0_t0
loss_patch1 = -5 * p1_t0
else:
gammas = [gamma0, gamma1]
loss_patch0 = -10 * FL(key_p0_t0, 1, gammas) * p0_t0 + 5 * FL(key_p0_t1, 0, gammas) * p0_t1
loss_patch1 = -10 * FL(key_p1_t0, 1, gammas) * p1_t0 + 5 * FL(key_p1_t1, 0, gammas) * p1_t1
# Back Prop
loss_patch0.backward()
loss_patch1.backward()
P0 = [normalized_p0_t0.item(), normalized_p1_t0.item()]
P1 = [normalized_p0_t1.item(), normalized_p1_t1.item()]
SCORES = [single_patch_out[i][0, map_idxs[i]].item() for i in range(patch_num)]
keys = key_p0_t0, key_p0_t1, key_p1_t0, key_p1_t1
return P0, P1, combined_patch_pic0, SCORES, patch_positions, random_bg, keys